Apple's solution to this seems pretty good, here's a snippet of the transcript from a WWDC 2020 talk:
> It might be tempting to think that reduced accuracy locations are simply your true location with some noise added in.
> But that is not the case. Simply adding some random noise on top of the user's true location is actually not that secure against someone who might wish to disambiguate the true location. Instead it's better to think of what happens under the hood as a quantization of sorts. Let me explain further.
> Let's look at an example area. Consider different users going on a drive in the vicinity of El Paso, Texas and Juarez, Mexico. Note that this is near an international boundary, with the US being in the northern half of the map, and Mexico being in the southern half, bordered by this pink line you see here. The true position of the user is represented by the cars while the reported position, that is to say what an app would see if it had reduced accuracy, will be shown by a circle. Again as I mentioned earlier in this presentation, these reported locations will contain the user's true location.
> As these users drive down Interstate 10 the location snaps to within different quantize regions of varying sizes. Both the red and yellow users are going to be quantized or "snapped" into the same regions as they drive along the road.
> But the amount of snapping will vary. In denser urban areas, this amount of quantization can be a couple of kilometers, but in less dense places like rural areas, it can be 10 kilometers or a bit more. A typical value for this quantization radius is around five kilometers. Note that despite the cars continuously driving along the road the region to which the true position quantities will not continuously update. This is because Reduced Accuracy locations are recomputed about 4 times per hour. So depending on exactly how fast these cars are moving there might be some lag between quantized snaps. The actual semantics of the quantization are intended to replicate what the user would expect to hear if they were to ask for an approximate answer to the question "Where am I"? We tried to ensure that the resultant reduced accuracy location is still going to be somewhere that the user actually expects to be. We use a location we have for you to place your approximate location on the appropriate side of the border here.
> Again though remember that the center point does not represent the location of the user. It only represents the center of a region where they approximately are. Approximate location snapping however might put you in a neighboring city especially if there are cities close together. To reiterate the bit about quantization around borders, let's see what happens to someone driving on the other side of the border in Juarez and what they might expect to see.
> Note that the blue car is quantized into regions whose centers are on the Mexican side of the border.
> > The actual semantics of the quantization are intended to replicate what the user would expect to hear if they were to ask for an approximate answer to the question "Where am I"?
> It might be tempting to think that reduced accuracy locations are simply your true location with some noise added in.
> But that is not the case. Simply adding some random noise on top of the user's true location is actually not that secure against someone who might wish to disambiguate the true location. Instead it's better to think of what happens under the hood as a quantization of sorts. Let me explain further.
> Let's look at an example area. Consider different users going on a drive in the vicinity of El Paso, Texas and Juarez, Mexico. Note that this is near an international boundary, with the US being in the northern half of the map, and Mexico being in the southern half, bordered by this pink line you see here. The true position of the user is represented by the cars while the reported position, that is to say what an app would see if it had reduced accuracy, will be shown by a circle. Again as I mentioned earlier in this presentation, these reported locations will contain the user's true location.
> As these users drive down Interstate 10 the location snaps to within different quantize regions of varying sizes. Both the red and yellow users are going to be quantized or "snapped" into the same regions as they drive along the road.
> But the amount of snapping will vary. In denser urban areas, this amount of quantization can be a couple of kilometers, but in less dense places like rural areas, it can be 10 kilometers or a bit more. A typical value for this quantization radius is around five kilometers. Note that despite the cars continuously driving along the road the region to which the true position quantities will not continuously update. This is because Reduced Accuracy locations are recomputed about 4 times per hour. So depending on exactly how fast these cars are moving there might be some lag between quantized snaps. The actual semantics of the quantization are intended to replicate what the user would expect to hear if they were to ask for an approximate answer to the question "Where am I"? We tried to ensure that the resultant reduced accuracy location is still going to be somewhere that the user actually expects to be. We use a location we have for you to place your approximate location on the appropriate side of the border here.
> Again though remember that the center point does not represent the location of the user. It only represents the center of a region where they approximately are. Approximate location snapping however might put you in a neighboring city especially if there are cities close together. To reiterate the bit about quantization around borders, let's see what happens to someone driving on the other side of the border in Juarez and what they might expect to see.
> Note that the blue car is quantized into regions whose centers are on the Mexican side of the border.
https://developer.apple.com/videos/play/wwdc2020/10660/